A user asked Heath for the bearing and distance between two parcels. Heath had no tool for it, so it said so instead of guessing. One "not helpful" tap fed a private, redacted feedback queue; a senior AI model read the queue, diagnosed a missing capability, and proposed a fix -- with an explicit warning not to let Heath invent bearings. An engineer reviewed it, built the real computation, added a regression test, and shipped. Heath now answers with a real bearing and distance -- and still declines when it genuinely can't compute one. That's the whole loop: feedback in, reviewed improvement out, grounding intact.
A question Heath wouldn’t answer
A right-of-way agent working a transmission corridor asked Heath, LandLedger’s assistant, a normal surveying question: the bearing and distance between two parcels.
Heath didn’t answer it. It said, plainly, that it didn’t have bearing or distance between specific parcels on hand – and offered what it could do instead: parcel dimensions, route crossings, alignment details.
That non-answer is the part worth slowing down on. Heath had no coordinate-geometry tool for a parcel-to-parcel bearing, so it declined. It did not produce a confident, plausible-looking “N 41 degrees E, 1,300 feet” pulled from nowhere. For anyone who has watched modern software state a bearing it never actually computed, the difference matters: a wrong bearing in a right-of-way file isn’t a glitch, it’s a liability. Heath is built so that every figure it states is computed from the records and reported – never guessed. When it can’t compute something, the correct answer is to say so.
So Heath was right. But the agent was also right that this is a question the product should be able to answer. Both things were true at once – which is exactly the situation a good feedback loop is for.
The tap that started it
Under that answer was a small “not helpful” control. The agent tapped it and left a one-line reason: Heath has the tools for this; it didn’t use them.
That tap didn’t email anyone or open a ticket. It wrote a single row to a private feedback log inside the customer’s own tenant – the question, the answer, the reason. Before any of it can be analyzed, it’s redacted: owner names and parcel IDs are replaced with consistent stand-ins (“Owner A”, “Parcel B”), dollar terms are masked, and the real figures that aren’t personal – acreages, distances – are kept so the gap stays diagnosable. The landowner’s name never leaves the boundary.
A downvote, in other words, isn’t a complaint that evaporates. It’s a structured signal with exactly enough context to act on, and nothing it shouldn’t carry.
What a senior model saw
On a regular cadence, a senior AI model reads the redacted backlog – not to answer users, but to critique the assistant. It gets Heath’s current instructions and its full tool list alongside the flagged cases, and it’s asked one thing: find the gap, and propose the smallest fix.
Here it was direct. The diagnosis: a genuine missing capability. Heath’s spatial tools could tell you which parcels a route crosses and how far a parcel sits from the line, but nothing computed a bearing and distance between two parcels. Heath’s decline was honest; the product was simply expected to do more.
The most important part of the proposal wasn’t the feature – it was the guardrail the model attached to it, unprompted:
Only add this once the tool genuinely computes the bearing from real coordinates. Never add wording that invites Heath to state a bearing it didn’t actually calculate, or you trade an honest “I can’t” for a confident wrong number.
That is the rule the whole product is built on, coming back out of the analysis on its own. The fix had to make Heath compute the answer, not merely sound like it could.
From flag to feature – reviewed
Nothing the model proposes ships on its own. A StrataLogic engineer reviewed the proposal, then built it for real: a coordinate-geometry routine that takes two parcels, computes the true bearing between them as a surveyor’s quadrant bearing and the distance in feet – straight from the parcel geometry in the customer’s records, projected to the local grid. Real coordinates in, real bearing out. A regression test went in beside it, pinning the exact behavior so it can’t quietly regress later. Then it shipped.
The same question now comes back with a real answer – a bearing like “N 44.9 degrees W” and a distance of roughly 1,485 feet – computed, not conjured. And when the two parcels can’t be pinned down (say, “the two parcels in that county” when the county holds a dozen), Heath asks which two rather than guessing. It got a new capability and kept its honesty.
Start to finish – flag, diagnosis, reviewed build, ship – took days, not a release cycle.
Why this is the point
It would be easy to frame this as “the AI improved itself.” It didn’t, and the distinction is the whole value. A different, stronger model proposed the change; a person reviewed and approved it; the improvement was a real tool computing real numbers, with a test to keep it honest. No weights were retrained, nothing changed in the dark.
That’s the loop we’re building Heath and Theo around: feedback-driven, human-reviewed, and grounded. The assistant that tells you “I don’t have that” is the one you can trust with the record – and a disciplined loop turns each of those honest gaps into a real capability, with someone accountable at every step. It doesn’t get better by magic. It gets better because your team’s real work points it at what matters, and because a human signs off before anything reaches you.
The more your team uses it, the sharper it gets – and you can always see exactly why.